840 karma · joined January 15, 2020
It doesn't matter that China/Whichever state actor is snooping on all your user's data. Either no-one finds out and you're good. Or the blast radius is _so_ wide, that all blame falls on Microsoft
However the risk from one dose is minimal. If your risk threshold is so close to 0 then we should be banning way more things today.
I'm not denying that people will self-harm. But the argument that people will self-harm with LSD needs to ignore that there's plenty of better drugs legalised already
Your average American that "can't handle cheese responsibly" will try it once, have a terrible 12h and swear off drugs for life. With no long term health effects from one dose.
I don't have strong opinions either way in this. But it's usually something that sparks a big flaming thread.
Obviously everyone is focussing on the AI part.
EDIT: Sorry scratch that. I've confused the URL. I didn't see this is a CNN article. This one https://www.bbc.co.uk/news/business-65631168 from bbc is much better (which I had read earlier). It shows the breakdown of where the jobs lost come from.
Which one do you think is more important to convince?
I know I'll sound kinda defeatist, but I just don't think it's possible to reliably setup a bureaucracy that's competent in funding innovative research. The incentives are just not aligned.
If you are truly investing in innovative companies. You'll inevitably spend taxpayer money on truly stupid stuff. The first time a major newspaper reports something like: "10 Million Euros spent on using bath sponges to desalinate water while people are living in the streets!" it's game over. The agency will be permanently risk averse.
It looks you just pattern-matched on the word _think_ and replied with a pre-made opinion about how AIs can't think. Ironic...
pip with wheels doesn't deal with non-python packages. I used to be in a horrible locked down corpo laptop. Conda was invaluable in getting stuff to run, like chromedriver, etc.
The theory is extremely interesting though. And better yet, it's falsifiable! If someone went around compared an RLHF model vs non-RLHF and found them equally likely to 'Waluigi' then we'd know this is false. And conversely if we found the RLHF more likely to Waluigi then it's evidence in favour.
The asymmetry in the hypothesis is really nice too. If this was true then I'd expect it to be possible to flip the sign in the RLHF step, effectively training it in favour of 'bad' behaviour. Then forcefully inducing 'Waluigi collapse' before opening to the public!
- There is some behaviour that we want the model to show, and the inverse we do not want it to. - Both are learned in the massive training phase - OpenAI used RLHF to suppress undesired behaviour, but it was ineffective because we have orders of magnitude less RLHF data.
That would imply that RLHF would slightly suppress the 'bad' behaviour, but it still would be easy to output it.
This is disproved by what the post is trying to explain: We see _increased_ bad behaviour by using RLHF. The post agrees with the premise that both good (wanted) and bad (unwanted) behaviour is learned during training. But it's proposing the 'Waluigi effect' to explain why RLHF actually backfires.
Now, tbh it does rely on the assumption that we are actually seeing more undesired behaviour than before. If that was false then it would falsify the Waluigi hypothesis.
By those standards every single human to ever live has been in prison. Humanity will only be free once we're in a post-scarcity society.
Now we're just seeing the good companies slimming down and bad ones failing.
I have quite a similar journey working in England, so it looks typical.
I used on an analysis that was taking long and it went way quicker.
I used it a month or two ago, so I don't remember specifics, but. The docs weren't unclear, they were bare compared to pandas tough. Once I found an example of how to do something I managed to do it. But a couple of times I opened a page and there was WIP message.
There's definitely a bit of unfairness in the comparison, as pandas as you say has been the standard for python for years. But that means I already know all the syntax and how the API works, so even just a method name or required params will get me a long way.
When I was fumbling around with polars I would have liked more examples of how to use each method/function. I mean, once the language clicks I wouldn't need it anymore. But at the start I really need that hand-holding. I'd have to reread other pages to remember simple syntax.
It's fast! The API is nice. But the documentation is not great. While using pandas I can just open the docs and find all functions and examples of how to use, etc. Polars felt really bare.
It's something I'd like to use in personal projects. But other data scientists would not accept using polars as it is today.
There was a blog written by Scott Alexander called 'Slate Star Codex', which was read by a lot of big tech entrepreneurs. Last year the New York Times wrote a piece on it, and told him they would reveal his name.
Scott took issue with being doxxed (too much to unpack in this, but caused a bit of a stir including here in hackernews) and took down the blog in protest. The blog had regular open threads for discussion, with some quite active regulars in there. datasecretslox was created for some of those to keep that space for discussion.
Lately Scott moved to substack and writes at AstralCodexTen.
It's the same as proofreading. Your brain just papers over repeated words or typos, or off-by-one errors.
Obviously there's a middle ground and it's unreasonable to ask someone to run every single line of code, but it's good to be aware of that.
If your point is that you can't spend money on illegal things, bitcoin is not illegal.
Also, who complained about immigrants? Where did this come from. And even if immigrants/farming had any relevance to the discussion co2 is likely to increase crop yields.
People mostly get cancelled by committing the latest heresy.